Speaker verification by combining information from magnitude and phase spectrum

The task of developing automatic speaker verification (ASV) system for security application is of considerable importance. This paper aims at developing a fusion strategy which combines both magnitude and phase information of the speech signal which yields a better performance when compared to conventional individual features. This paper employs Mel frequency cepstral coefficients (MFCC) and modified group delay function (MODGDF) as feature extraction techniques and Artificial Neural Network (ANN) as the speaker modeling method to perform ASV. TIMIT database is used for the proposed study. The paper also compares the performance of different fusion strategies by computing Receiver operating characteristic (ROC) for evaluation of the system performance.